一种用于输电线路绝缘子多缺陷检测的轻量级网络

A Lightweight Network for Multi-defect Detection of Transmission Line Insulators

针对绝缘子多类型缺陷检测速度慢、检测精度低的问题,提出一种用于输电线路绝缘子多缺陷检测的轻量级网络(Multi-Defect Detection Network, MDDNet),该算法主要针对绝缘子电弧烧伤和绝缘子伞裙破损的多类型绝缘子缺陷联合检测。基于Ghost-C2f模块构建GC-Darknet53特征提取网络,增强特征提取能力并较少特征冗余;引入三尺度融合(Tri-Fusion)机制构建新型TF-Neck颈部网络,充分融合深层语义信息与浅层的细粒度信息,提高小目标缺陷检测精度;选用结构相似性交并比(Structural Similarity Intersection over Union, SIoU)损失函数提高模型定位能力。实验结果表明,提出的MDDNet模型平均精度均值(mean Average Precision mAP)达到92.1%,与YOLOv5相比,在参数量减少了20%的情况下mAP提升了3.0%,与其他现有一阶段算法相比,MDDNet算法检测速度达到86.1帧/秒,能够在保证轻量化的同时提高检测精度,满足绝缘子多缺陷检测的应用需求。

To solve the problems of slow detection speed and low detection accuracy of multi-type defects of insulators,a Multi-Defect Detection Network(MDDNet)lightweight network for multi-defect detection of transmission line insulators is proposed.This algorithm mainly targets at the arc burns and shed breakage of insulators for joint detection of multi-type insulator defects.Firstly,the GC-Darknet53 feature extraction network is constructed based on the Ghost-C2f module to enhance the ability of feature extraction and reduce feature redundancy.Then,a Three-scale Fusion(Tri-Fusion)mechanism is introduced to construct a new TF-Neck neck network,which fully integrates the deep semantic information with the shallow fine-grained information to improve the accuracy of defect detection for small targets.Finally,the Structural Similarity Intersection over Union(SIoU)loss function is chosen to improve localization ability of the model.Experimental results show that the mean Average Precision(mAP)of the proposed MDDNet model reaches 92.1%.Compared with YOLOv5,the mAP improves by 3.0%with a 20%reduction in the amount of parameters.Compared with other existing one-stage algorithms,the MDDNet algorithm achieves a detection speed of 86.1 frame/s,which can improve detection precision and ensure lightweight,satisfying the application requirements of insulator multi-defect detection.